Airbnb’s customer support cost per booking fell more than 17% year over year. CEO Brian Chesky said on the Q3 2026 earnings call that the drop was driven in large part by improvements to its AI assistant. CX Dive reported it on 10 November 2026.

The article adds a few details. The assistant launched about a year and a half before the call. In Q3 it resolved approximately 40% of all customer support issues without human agents. Chesky said customer service NPS is “as high as it’s ever been” and that the improvements are “just the beginning.” Airbnb is also adding voice support in the US.

The excitement is about the 17%, but the metric is the lesson

Most coverage will focus on the percentage. I think the more useful part is what Airbnb divided by. It chose cost per booking.

Bookings are the number Airbnb already runs on. Support spend gets measured against the thing the business exists to produce. If bookings grow and support cost grows slower, the AI is doing something real. If the AI deflects a lot of tickets but guests stay stuck, cost per booking will not fall the way Chesky described. The metric has to be moved by actual work.

Chesky also reported NPS in the same breath as the savings. The article doesn’t say how Airbnb weighs the two. But reporting both tells you the cost figure was not allowed to improve at the expense of the guest experience.

Counting deflected tickets rewards the wrong thing

Many businesses measure AI support by tickets deflected or chats contained. Those numbers go up when the AI gets in the customer’s way. A customer who gives up and leaves counts as a deflection. A customer who asks the same question three times counts as three contained chats.

Resolving beats deflecting. Success means the customer’s problem got solved. A cost-per-unit metric forces that, because an unresolved problem comes back as another contact, a refund request or a lost repeat order.

What this looks like for a normal operator

Take a few examples:

  • A shop could track support cost per order. If the AI answers “where’s my order” by checking the order and replying, cost per order drops without the customer being sent to a form.
  • A clinic could track front-desk cost per booked appointment. Messages that end in a booking count. Messages that end in “please call us” don’t.
  • A service business could track support cost per completed job, and watch repeat contacts on the same job as a guardrail.

In each case, pick the number your business already reports. Then add one check that the AI cannot game, such as repeat contacts or customer satisfaction. The AI has to move the first number without hurting the second.

40% is an early number, and I would not treat it as the ceiling

I read the 40% as a stage. The assistant is about eighteen months old, and Chesky himself calls the gains the beginning. The article doesn’t say what mix of issues makes up that 40%, so I can’t tell you how much of the rest is hard and how much is simply not built yet.

My own view is that a competent AI can already handle most transactional and operational enquiries. I would put it at 80 to 90%, and that is my opinion of what competent AI can do in general. It isn’t a figure for any one product. Order status, bookings, changes, refunds within a set limit and basic policy questions are routine work. People are still needed for the cases that go beyond the limits and for the moments where a human conversation sets a business apart, such as a big-ticket purchase.

Airbnb is a large marketplace with its own systems and 20 years of history, so its starting point differs from a small shop’s. But the direction applies to small businesses too. A competitor running AI on its transactional enquiries will serve more customers at lower cost than a team answering every message by hand.

Where this connects to what I build

This is the reason AeroChat is built around workflows that act in a merchant’s systems, such as checking an order or booking a slot. Answering a question doesn’t change cost per booking. Completing the task in the conversation does.

What a business should do

  1. Choose one metric your business already tracks, such as cost per order, cost per booking or cost per job.
  2. Add a guardrail the AI can’t game, such as repeat contacts on the same issue or a satisfaction score.
  3. Record the baseline before you switch anything on, so the comparison is honest.
  4. Review the conversations the AI could not finish. Each one shows a workflow to build next.
  5. Decide where a person adds value to the customer, and keep people there.

Source: CX Dive: Airbnb reduces customer support cost per booking 17%